JPH0815261B2 - Adaptive transform vector quantization coding method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for compressing the information of sampled audio signals whose frequency of occurrence is biased.

[0002]

2. Description of the Related Art It is important to highly efficiently encode a voice or audio signal for transmission or recording, and many examples have been reported. Among them, the following are examples of vector quantization, adaptive transform coding, or a combination thereof. As a mid-band (4.8 to 9.6 kbps) audio coding method for transmission and storage applications, there are adaptively vector-quantized linear prediction residual signals in the frequency domain and scalar-quantized ones. A method of adaptive transform vector coding that combines the following: (The Institute of Electronics and Communication Engineers, Journal of Japan, '84 / 10 Vol. J67-ANo. 10, Moriya,
HONDA), Low bit rate of voice (2.4 to 1.7 Kbp)
s) As a coding method, a fuzzy-vector quantization is simplified, and a complementary vector quantization method using two vectors and their linear combination coefficients (IEICE Information Theory Research Group '91 / 2 IT90-103) , Asakawa and others).

All of these methods apply vector quantization. The first method is to perform DFT (discrete Fourier transform) on a linear prediction residual signal and divide its output spectrum into a plurality of bands. , Adaptively allocating bits to each band and vector quantizing, the second one is
PSE analysis of the voice signal, vector quantization of the impulse response waveform obtained from this, and one vector of the code vector closest to the input vector and a plurality of previously selected candidate vectors close to this code vector One vector with the smallest distortion is selected from the linear combination of.

In the first method, since the variance of the prediction residual spectrum can be regarded as being substantially flat, the bands are equally divided, and the number of bits assigned to each band is adaptively calculated from the average power of each spectrum. However, the number of bits b allocated to each band is determined by a predetermined transfer rate per bit.
If the number of bits is less than or equal to the number of bits h calculated from the multiplication of the spectrum number in the band and the number of spectrums in the band, vector quantization is performed as it is. With the candidates left, as a second step, a (b-h) -bit scalar-quantizer is provided around each centroid to determine the code with the minimum distortion. But,
In this method, the prediction residual is used, and it is difficult to apply it as it is to a signal with a large change such as audio.

The second method is fuzzy-vector quantization, in which many code vectors can be combined as a result even though the code vector has a small dimension, but many code vectors and coupling coefficients are transferred. Since it is necessary, two vectors, which are the ultimate form of fuzzy-vector quantization, and the linear combination coefficient connecting them are used. A conventional example will be described with reference to FIG. The audio signal applied to the input terminal 90 is input to the PSE analyzer 91 and the pitch extractor 92, and the envelope of the power spectrum is calculated in relation to the pitch. The impulse response converted from this spectral envelope has its amplitude normalized by the maximum value, samples of only the dimensions of the vector are taken out, vectorized by the vectorizer 93, and the complementary vector quantizer 9
4. In the code book 95, two vectors and one vector combination are quantized as a linear combination coefficient. These data are multiplexed by the time division multiplexer 96 and transferred. During reproduction, these data are respectively distributed by the distributor 97, and the complementary vector quantizer 98 and the codebook 99 are distributed.
And is extrapolated by linear prediction so that the LPC extrapolator 100 removes the effect of truncation of the impulse waveform in encoding. Finally, the PSE synthesizer 101 synthesizes the waveforms by superimposing the pitch period intervals of the impulse response, and a decoded speech signal is obtained from the output terminal 102.
This method basically targets speech signals and vector-quantizes the pitch impulse response obtained from PSE analysis, but cannot be applied to audio signals having different spectral structures.

[0006]

As described above, the vector quantization is an efficient quantization method utilizing the bias of the statistical distribution of data, and if the codebook size is increased,
Quantization distortion can be reduced, but problems occur in the storage capacity of the codebook, the amount of search processing, and the like. Further, the fuzzy-vector quantization can reduce the quantization distortion by using the same codebook as compared with the normal vector quantization, but it is difficult to apply it to the encoding as it is because the amount of information such as a class function increases. Is. Therefore, it has been attempted to represent only two vectors and one linear combination coefficient as the ultimate form of a fuzzy vector, but when a wide band audio signal is encoded at a low transfer rate, the spectrum At a low average coding rate, there is a problem that sufficient bits cannot be assigned to the vector and the linear combination coefficient, which rather increases the distortion.

It is an object of the present invention to solve the above-mentioned drawbacks and to provide a coding method with less distortion by increasing the quantization efficiency by combining fuzzy-vector quantization and conventional vector quantization.

[0008]

According to the present invention, in order to achieve this object, the level of an input signal waveform is normalized by dispersion and then converted from a time domain to a frequency domain, and a transformed spectrum is divided into a plurality of bands. Obtaining the average power of each output, normalizing each of the conversion spectra by this average power, estimating the size of each individual spectrum by interpolation based on the average power, thereby the quantization width of each spectrum and the quantum to be assigned. The number of coded bits is calculated, the number of spectra in each band is used as the number of dimensions of the vector, and the number of levels obtained by multiplying this by the average spectrum coding rate of each band is used as the vector quantization, and two vectors and Fuzzy-vector quantization represented by one linear combination coefficient connecting
The one with smaller distortion is selected.

Also, in the above fuzzy-vector quantization, the conventional vector quantization is applied only when a sufficient number of bits are not allocated to two vectors.

Further, the average spectrum coding rate of each band
In the large band, the band is subdivided so that the dimension of the vector becomes smaller.

In the low frequency band, the vector is divided so that the dimension of the vector becomes small.

[0012]

In order to lower the transfer rate, a vector is represented by only two vectors, which are the ultimate form of fuzzy-vector quantization, and a linear coupling coefficient connecting the two vectors, and a wide band audio signal is transferred at a low rate. When encoding with a rate, in the low band of the average coding rate of the spectrum due to the limitation of the transfer rate, it is not possible to allocate sufficient bits to the vector and the linear combination coefficient, rather increasing the distortion. The problem of being a normal vector quantization and two vectors and fuzzy-vector quantization represented by a linear combination coefficient are used together,
The distortion is reduced by selecting the one with the smaller distortion.

Further, in order to reduce the amount of calculation, the value obtained by multiplying the average spectrum coding rate of each band, which is back-calculated from the transfer rate in the fuzzy-vector quantization, by the number of spectra is 2 Only when a sufficient number of bits are not allocated to one vector, normal vector quantization is applied to reduce distortion.

Further, once the spectral band of the block is evenly divided, and when the number of bits assigned to each band is larger than the predetermined number of bits, the band is further divided into two parts, and the average power of the band is divided. By recalculating the number of allocated bits, the number of vector dimensions is suppressed and the codebook size of each band is adjusted in the band with high average power.
At the same time, the codebook size is prevented from becoming large, the average power calculation accuracy is improved, and the coding accuracy of each spectrum is improved.

Further, in a normal audio signal, the average power is higher at a low frequency band than at a high frequency band, so that a predetermined low frequency region is redivided by a predetermined number of divisions to obtain a code. It is intended to improve the conversion accuracy.

[0016]

EXAMPLES Specific examples will be described in detail below. FIG. 1 is a block diagram of an encoder according to a first embodiment to which the adaptive transform vector quantization encoding method of the present invention can be applied. The encoding procedure will be described with reference to FIG.

A broadband audio signal is applied to the input terminal 10 and temporarily stored in the buffer 11. Buffer 1
A block composed of a plurality of samples is cut out from the signal in 1, the variance is obtained by the variance calculator 12 based on (Equation 1), and the quantizer 13 quantizes and inputs the variance to the multiplexing circuit 30.

[0018]

[Equation 1]

The signal of the clipped block is the inverse quantizer 1 in which the quantized dispersion value is the same as that on the reproducing side.
4 and thereby normalized by the variable gain amplifier 15. The normalized signal is transformed from the time domain to the frequency domain as shown in FIG. 5A by the orthogonal transformation circuit 16, for example, DCT. Here, x (k) is a conversion coefficient. The DCT coefficient (spectrum) is divided into a plurality of bands, and the average power in that band is calculated by the in-band average power calculator 17 according to (Equation 2), as shown in (FIG. 5B). . Here, p (l) is the average power of each band.

[0020]

[Equation 2]

The average power thus obtained is coded by the quantizer 18 and input to the multiplexing circuit 30. This quantized average power is decoded by the inverse quantizer 19,
Linear interpolation is performed by the interpolator 20. The coefficient size σ _K (variance) is estimated based on this, and the normalized gain calculator 21 and the bit allocator 22 allocate adaptive bits to the DCT coefficients.

Here, the adaptive bit allocation to each band is calculated based on (Equation 3).

[0023]

(Equation 3)

Thereafter, the normalized coefficient is input to the vector quantizer 26 and the fuzzy-vector quantizer 24. In the fuzzy-vector quantizer 24, the vector closest to the spectrum of a certain band is represented by two vectors and one linear coupling coefficient connecting them according to the following procedure. First, the code vector closest to the spectrum of a certain band is selected, and then based on (Equation 4), a plurality of candidate code vector groups close to this preselected code vector are selected. One vector is fetched.

[0025]

[Equation 4]

[0026] Here, y ^ is fuzzy - vector, v ₁ is the closest co - a Dobekutoru, v ₂ candidate vectors, m is the coupling coefficient. Here, the coupling coefficient m for linearly coupling v ₁ and v ₂ is changed, and the distortion is calculated by the distortion calculator 25.
If the distortion increases beyond the single vector v ₁ , the candidate vector is discarded. When the distortion is small, the distortion is stored, and the distortion is similarly calculated for the second candidate vector. In this way, one code vector v ₂ with the smallest distortion is selected. The distortion of the vectors in these candidate code vector groups does not always decrease depending on the positional relationship with the nearest code vector, but may increase on the contrary. This is shown in FIG. In FIG. 6, (a) shows the case where the strain decreases, and (b) shows the case where the strain increases.

In the conventional vector quantizer 26, the code vector closest to the spectrum of the band is selected from the code book represented by the number of bits allocated to the band. The distortion of the output of the vector quantizer 26 is calculated by the distortion calculator 27 and the fuzzy-vector quantizer 2
The smallest distortion of 4 is compared with the distortion comparator 28.
Of these, the one with less distortion is selected by the selection circuit 29 and input to the multiplexing circuit 30. In the multiplexing circuit 30, the finally selected vector and the coupling coefficient (in some cases only vector), the selection signal of the normal vector or the fuzzy-vector, the quantized in-band average power, and the quantized dispersion information are output. Time division multiplexed. Compressed data
Data is sent to the output terminal 31. Here, in each vector quantizer, a number obtained by multiplying the average spectrum coding rate of each band by the number of spectra is calculated by the level number calculator 32, and vectorized based on this.

Next, a second embodiment of the present invention will be described with reference to FIG. In (FIG. 2), the same components as those in (FIG. 1) are designated by the same reference numerals. Only the points different from the first embodiment will be described below. In the first embodiment, the distortions of the normal vector and the fuzzy-vector are calculated in all the bands in one block, and the smaller one is calculated, but in the second embodiment, the fuzzy-vector quantum is calculated. As shown in (Fig. 2),
Based on the number of bits assigned to each coefficient obtained by linear interpolation from the average power of each band, the number of levels calculated by multiplying the average spectrum coding rate of that band by the number of spectra of that band 32, and from the number of levels (number of bits) assigned to each band in the block,
The bit number determiner 33 subtracts the number of bits of the coupling coefficient m having a predetermined precision from the remaining (v ₁ +
The normal vector quantization is adopted only when the number of bits of v ₂ ) is less than the predetermined number of bits of two vectors. For simplification, the number of bits of (v ₁ + v ₂ ) allocated to each band in the block is
Normal vector quantization is adopted only when the number of bits of the predetermined coupling coefficient m is not more than twice.

A third embodiment of the present invention will be described with reference to FIG. The same blocks as those in (FIG. 2) are designated by the same reference numerals, and detailed description thereof will be omitted.

First, the spectral band of the block is equally divided, and adaptive bits are allocated in the same manner as in normal adaptive transform coding. If the number of bits at this time is calculated by the level number calculator 32 and the number of bits determiner 33 determines that the number of bits is larger than the predetermined number of bits, the number of bits determiner 33 performs the orthogonal transformation on the re-division instruction signal. It is sent to the unequal divider 34 which is connected directly to the circuit 16, whereby the band is further divided into two equal parts, the average power of the band is recalculated, and the number of allocated bits is again calculated based on it. The subsequent processing is the same as that of the second embodiment. This is the average transfer rate for each band.
, And if the average power is large, re-divide again, reduce the number of vector dimensions in the band with high average power to make the codebook size of each band uniform, and avoid increasing the codebook size. Further, the accuracy of estimation of each spectrum is improved by improving the accuracy of calculation of average power. Here, the switching of the bands in the descending order of average power is not performed because the transfer information of this switching information increases. It is also possible to apply this processing to the first embodiment.

Next, a fourth embodiment of the present invention will be described with reference to FIG. The same blocks as those in (FIG. 3) are designated by the same reference numerals. Since a normal audio signal has a higher average power in a low frequency band than in a high frequency band, a predetermined low frequency region is subdivided by the unequal band divider 34 by a predetermined division number. It This is also the same purpose as the third embodiment.

In the decoding of each embodiment, basically the reverse processing of the encoding is executed. As an example, a block diagram of the decoding process of the first embodiment is shown in FIG. Since the operation and contents are clear, the description is omitted. In the decoding for the encoding process of (FIG. 2) to (FIG. 4), the bit number determiner is added to (FIG. 7), and the transfer of the selection signal becomes unnecessary.

By the encoding method as described above, an audio signal having a wide band is low, for example, 128 kbps.
High-quality coding can be performed with the following transfer rate.

[0034]

As described above, according to the present invention, adaptive quantization width and adaptive bit allocation are performed on orthogonally transformed coefficients, normal vector quantization is performed, and two vectors and a combination coefficient for linearly combining them. By adaptively combining the fuzzy-vector quantization represented by, the distortion can be suppressed and the encoding can be performed at a low transfer rate. Also, the spectral band of the orthogonally transformed coefficients is adaptively divided,
The audio signal can be encoded with high quality while keeping the codebook size of each band uniform and suppressing the storage capacity of the codebook small.

[Brief description of drawings]

FIG. 1 is a block diagram of an encoding process according to a first embodiment of this invention.

FIG. 2 is a block diagram of a coding process according to a second embodiment of the present invention.

FIG. 3 is a block diagram of an encoding process according to a third embodiment of the present invention.

FIG. 4 is a block diagram of an encoding process according to a fourth embodiment of the present invention.

5A is an explanatory diagram of coefficients (spectrum) converted from a time domain to a frequency domain by orthogonal transform (DCT) processed in a conventional example or an embodiment of the present invention. FIG. (B) In encoding and decoding, it is explanatory drawing of the average power of each band as auxiliary information for quantizing or dequantizing the coefficient transformed by orthogonal transformation (DCT). (C) It is explanatory drawing of the magnitude | size of each spectrum estimated from the average electric power.

FIG. 6A is an explanatory diagram of a relation between vector arrangement and distortion when distortion is reduced in two vectors in a fuzzy vector in which two vectors are linearly combined. (B) Two vectors and fuzzy linear combination of them
It is an explanatory view of the relation between vector arrangement and distortion when distortion increases in two vectors in a vector.

FIG. 7 is a block diagram of a decoding process according to the first embodiment of this invention.

FIG. 8 is a block diagram of a coding process of a conventional adaptive transform coding.

[Explanation of symbols]

 10 Input Terminals 11 Buffer 12 Variance Calculator 13, 18 Quantizer 14, 19 Inverse Quantizer 15, 23 Variable Gain Amplifier 16 Orthogonal Transform Circuit 17 In-Band Average Power Calculator 20 Interpolator 21 Normalized Gain Calculator 22 bits Allocator 24 Fuzzy-vector quantizer 25, 27 Distortion calculator 26 Vector quantizer 28 Distortion comparator 29 Selection circuit 30 Multiplexing circuit 31 Output terminal 32 Level number calculator 33 Bit number determiner 34 Unequal divider 70 Input terminal 71 Distributor 72 Fuzzy-vector quantizer 73 Vectorizer 74 Selection circuit 75,81 Inverse quantizer 76 Interpolator 77 Normalization gain calculator 78 Bit allocator 79,82 Variable gain amplifier 80 Inverse orthogonal transformation circuit 83 buffer 84 output terminal 85 level number calculator

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Internal reference number FI Technical indication H04B 14/00 E

Claims (4)

[Claims] 1. The variance of an input voice or an audio signal is obtained, the input signal is normalized, then the time domain is transformed into the frequency domain, and the transformed spectrum is divided into a plurality of bands.
Obtaining the average power of each output, normalizing each of the conversion spectra by this average power, estimating the size of each individual spectrum by interpolation based on the average power, thereby the quantization width of each spectrum and the quantum to be assigned. In the adaptive transform coding method for calculating the number of coded bits, the number of spectra in each band is defined as the number of dimensions of the vector, and the number of levels obtained by multiplying this by the average spectrum coding rate of each band is vector-quantized, and An adaptive transform vector quantization coding method characterized by performing fuzzy-vector quantization represented by two linear vectors and one linear combination coefficient connecting them, and selecting the one having smaller distortion. 2. The variance of the input voice or audio signal is obtained, the input signal is normalized, then the time domain is transformed into the frequency domain, and the transformed spectrum is divided into a plurality of bands.
Obtaining the average power of each output, normalizing each of the conversion spectra by this average power, estimating the size of each individual spectrum by interpolation based on the average power, thereby the quantization width of each spectrum and the quantum to be assigned. In the adaptive transform coding method for calculating the number of coded bits, the number of spectra in each band is defined as the number of dimensions of the vector, and the number obtained by multiplying this by the average spectrum coding rate of each band is defined as the number of levels. An adaptive transform vector quantum characterized by performing fuzzy-vector quantization represented by one linear combination coefficient connecting them and applying conventional vector quantization when a sufficient number of bits is not allocated to two vectors. Coding method. 3. Obtaining the variance of an input voice or audio signal, normalizing the input signal, converting it from the time domain to the frequency domain, dividing the transformed spectrum into a plurality of bands,
Obtaining the average power of each output, normalizing each of the conversion spectra by this average power, estimating the size of each individual spectrum by interpolation based on the average power, thereby the quantization width of each spectrum and the quantum to be assigned. In the adaptive transform coding method for calculating the number of coded bits, vector quantization and fuzzy-vector quantization represented by one linear coefficient connecting the two vectors are applied to adaptively switch the coding method. An adaptive transform vector quantization coding method characterized in that the band is subdivided so that the number of dimensions of the vector becomes small in the band having a large average spectrum coding rate of each band. 4. Obtaining the variance of an input voice or audio signal, normalizing the input signal, converting it from the time domain to the frequency domain, dividing the transformed spectrum into a plurality of bands,
Obtaining the average power of each output, normalizing each of the conversion spectra by this average power, estimating the size of each individual spectrum by interpolation based on the average power, thereby the quantization width of each spectrum and the quantum to be assigned. In the adaptive transform coding method for calculating the number of quantization bits, vector quantization and 2
Applying fuzzy-vector quantization expressed by one vector and one linear coefficient connecting them, and adaptively switching coding method, the division is performed so that the dimensionality of the vector becomes small in the low frequency band. Characteristic adaptive transform vector quantization coding method.